Beyond Gain: Measuring Skill Internalization for Agent Skill Retirement
Yanze Wang, Pengfei Yao, Jun Sun, Chuanrui Hu, Yafeng Deng and colleagues (EverMind, Shanda Group, Peking University, Southeast University) show that retiring agent skills on the with-versus-without gain removes many skills that no model has actually internalized.
Ask this paper
Problem. A near-zero gain occurs both when a model has internalized a skill and when neither model nor skill delivers the capability, a state the authors call uncovered.
Method. Withhold the skill, score the model on a task synthesized from it, and trust a high unaided score only if another model gains from the skill; nearly half of high scores come from tasks no model benefits from.
Results. Under four models a published gain rule with one task per skill removes over half of the uncovered skills. On SkillsBench, 28 of 31 retirements rest on probes no run fully passed, while retiring on the unaided score costs no detectable reward.
Reliability. Unaided verdicts reproduce across tasks more often (Cohen's kappa 0.42 vs 0.23) but differ across models. The authors release 1,251 skills with tasks and scores under four models.
Abstract
LLM agents retrieve skills from public hubs that hold up to about 100,000 entries and serve many models. As models improve, a hub must find the skills a given model has internalized, meaning it can do without them, and retire them for that model, yet internalization has not been measured with known reliability. Published rules that compare runs with and without a skill retire on the gain alone. The gain is near zero both for an internalized skill and when neither the model nor the skill delivers the capability, a state we call uncovered. We instead withhold the skill, score the model on a task synthesized from it, and trust a high unaided score only when another model gains from the skill, since nearly half of high scores come from tasks no model benefits from. Under four models, a published gain rule applied with one task per skill removes over half of the uncovered skills. On SkillsBench, 28 of its 31 retirements rest on probes no run passed in full, while retiring on the unaided score costs no detectable reward. Unaided verdicts reproduce across tasks more often than the gain rule's (Cohen's kappa 0.42 against 0.23) but differ across models, so a hub must read them per model. We release 1,251 skills with tasks and item scores under four models.